What it is
Grass pollen has been largely overlooked in studies of how grasslands evolved, because the pollen of most grass species cannot be told apart with traditional optical microscopy. Using a semisupervised deep-learning strategy, the authors trained convolutional neural networks (deep-learning models for images) on superresolution images of known modern grass pollen and unlabeled fossil grass pollen. With the features the networks learned, they estimated taxonomic diversity and distinguished C3 from C4 grasses (two photosynthetic types), validated the method on modern pollen assemblages, and then applied it to a 25,000-year lake-sediment record from Mt. Kenya.
Why it matters
In the 25,000-year lake-sediment record from Mt. Kenya, the species diversity of these montane grasslands declined substantially between 21,500 and 16,000 years ago, coinciding with the most severe regional cooling of the last ice age, and the share of C4 grasses gradually decreased after the transition from the late glacial to the Holocene, associated with higher temperatures and decreasing fire activity. The shifts were set against atmospheric CO₂ and local records of temperature, precipitation and fire. The authors conclude that features of pollen morphology learned by the networks can advance pollen analysis and enable robust estimates of changes in grass diversity and C4 abundance in ancient grasslands.
Underlined numbers link to their source. Every metric and quoted figure is listed under Sources and data below.
Filed underGeology and Paleoclimatology Research, Plant Diversity and Evolution, Pleistocene-Era Hominins and Archaeology